Closed launch · by invitation
relaylet sits between your coding agent and the model provider. It forwards every byte unchanged and records every attempt honestly: what was sent, what it cost, and what is still unknown.
Works with the clients you already use: Codex (OpenAI Responses) and Claude Code (Anthropic Messages).
HOW IT WORKS
Your client's request and the provider's stream pass through unchanged. That includes the opaque reasoning state that tool loops depend on. Nothing is rewritten, no model is silently substituted, and nothing is retried behind your back.
Every attempt is recorded and given one of three states. Settled means it was charged from provider-reported usage. Not charged means nothing was processed. Unresolved means it was sent but the outcome is unknown. Unknown is never shown as zero.
Credit for a request's worst-case cost is held before the request leaves. If you bring your own provider key, it is encrypted at rest. Keys are scoped, rate-limited and revocable, and each one is shown only once.
WHAT YOUR AGENT'S OWN COUNTER LEAVES OUT
In one qualification run the provider's stream was cut mid-response. Codex then retried the request on its own, with the same content in a different JSON key order. Its summary printed tokens used 23,234, which counts only the completed requests.
relaylet's record for the same session showed three upstream requests that may have been billed. One of them had unknown usage, and it was the attempt Codex had retried.
attempts 3 | forwarded 3 | client retries 1 upstream requests that may be billed: 3 (usage reported 2, unknown 1) - #2 the upstream stream ended abruptly after 1530 bytes without a terminal event. The provider may still have processed and billed it; relaylet did not retry. - #3 completed. It repeats attempt #2 (same request content: a client retry).
Observed with the real Codex 0.160.0 client against relaylet's deterministic stand-in provider. No real model was involved.
EVIDENCE, NOT CLAIMS
Each client was run through eight scenarios: a two-turn tool loop, a cut stream, a rate limit, a request budget, the wrong model, the wrong key, and cancellation before the provider had sent anything. The table shows what was observed. The evidence level is stated plainly: real clients, scripted provider, not yet live.
| Codex 0.160.0 | Claude Code 2.1.288 | |
|---|---|---|
| Scenarios passed | 8 / 8 | 8 / 8 |
| Opaque reasoning state returned intact | yes (encrypted_content) | yes (thinking signature) |
| Rate limit (429) | does not retry | retries and recovers |
| Rejected key (401) | 5 retries | 11 attempts over ~3 minutes |
| Cancel before the provider responds | upstream closed in ~1 s | upstream closed in ~1 s |
| Identifiers sent to the provider | installation, session, thread, turn ids; working directory | device id, session id; local paths |
| Traffic outside the model path | chatgpt.com, ab.chatgpt.com, github.com, api.github.com | none observed with CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 |
Claude Code sends CLAUDE.md files from every parent directory of where it is launched. Launch it from your project, not from a folder that holds private notes.
SECURITY
STATUS · 2026-10
This page, served as static files.
The hosted gateway and the customer console are not open, and no payments are being taken. Reselling model access requires the providers' written permission. Until we have it, accounts will use your own provider key.